Research On Target Detection Algorithm Based On Deep Learning

The rapid development of world industry,Deep learning has been widely applied from nobody before,among which the research direction of target detection,The research direction of target detection is an important part of deep learning,has been paid more and more attention by the industry.As the detection environment is more and more complex and the detection equipment is more and more advanced,The field of target detection has developed from the traditional algorithm at the beginning to the current deep learning algorithm,in which there have been many researches on the detection of small targets and the detection speed This paper mainly conducts research and experiments on these two aspects,and the specific contents include the following:Firstly,the basic principles of fayer-rcnn and YOLO-v3 of deep learning target detection networks were analyzed,including their overall process of detecting targets,Target detection internal network structure analysis,feature extraction,loss function construction,and principle of lightweight network MobileNet-v3,which laid a theoretical foundation for subsequent algorithm optimization on this basis.Secondly,an improved Faster R-CNN small target detection algorithm is proposed.Aiming at the problem that Faster R-CNN internal network structure makes insufficient use of feature map information,a bottom-up reverse side connection path is added to the internal network structure to optimize the target detection method,and then ms-coco,a public data set,is used for training and testing.Experimental results show that compared with fFaster R-CNN network before improvement,the detection accuracy of bounding box and target is improved to a certain extent,especially for small and medium-sized targets.Finally,An improved YOLO-v3 target detection algorithm is presented in the initial YOLO-v3 network structure,the original backbone network Darknet-53 replaced by a more lightweight MobileNet-v3,set up the training parameters,training model on MSCOCO data set,and save the model,the performance testing model,Detection accuracy is rarely reduced as a prerequisite,improve the detection speed,and CPU usage in the operation to reduce the effect is significant.Then,the improved YOLO-v3 network was used to train the infrared face data and detect the infrared face.The model was deployed to FPGA7020 and the DPU was used for parallel acceleration.The experimental detectionreasoning time was obtained.It lays a foundation for real-time detection in mobile devices.In this paper,the algorithm of small and medium target detection algorithm and detection speed two classical directions are studied,the performance of the algorithm model is further improved,so that it can show a more ideal detection effect in a more complex environment to lay a foundation for identification.